基于 GPU 的后向投影 SAR 成像算法

摘要

Back-Proj ection(BP)is an accurate time domain algorithm for processing synthetic aperture radar(SAR)data.However,it is difficult to image in real time by its big computation load.Graphic process-ing unit(GPU),which involves the ability of powerful floating point arithmetic and the capacity of high paral-lel processing,provides a good platform for back-projection implementation.A parallel BP algorithm using GPU is presented in this paper,four optimization methods are applied,and the resolution to avoid bank con-flict problem of shared memory is proposed,which reduces the access time of shared memory.The results of simulation data demonstrate that the proposed algorithm is accurate and an acceleration ratio of 70 over a single-threaded CPU implementation is achieved.%后向投影（BP）是一种精确的时域合成孔径雷达（SAR）成像算法，但是其巨大的运算量很难满足实时成像的要求，图形处理器（GPU）具有强大的浮点运算和高度的并行处理能力，为 BP 算法的实时成像提供了一个很好的平台。提出基于 GPU 的并行化 BP 算法，利用了四种优化方法对并行化 BP 算法进行加速，并且针对共享存储器的 bank 冲突问题提出了相应的解决方法，减少了共享存储器访问时间。最后给出仿真数据的成像结果，结果表明，与传统的基于 CPU 单线程的 BP 算法相比，成像速度可达到70倍以上的提升。